Design of Gesture Vocabularies through Analysis of Recognizer Performance in Gesture Space

Abstract

This paper proposes a different paradigm for the design of gesture vocabularies that could be used, for instance, in a pen-gesture interface. In contrast to the usual approach in which a classifier is trained with multiple sketches drawn by an end-user who has a tentative set of sketch templates provided by the system designer, we intend to explore the gesture space to automate the design of the gesture vocabulary. Instead of training, a model of the gesture space of the recognizer is analyzed and a selforganized set of gestures with improved recognition rate is suggested for the designer or the end-user which then is able to filter and brush up the resulting vocabulary. As a result, simpler recognizers can be derived, vocabularies can be personalized by users without specialist knowledge of the behavior of the recognizer and a quasi-optimal vocabulary can be obtained regarding recognizer performance.